The people who curate training data use all kinds of tricks, yes. They also constantly try new ideas and see what helps. Btw, learning from your own generated

The people who curate training data use all kinds of tricks, yes. They also constantly try new ideas and see what helps. Btw, learning from your own generated data isn't necessarily bad, if you have an external source of validation. See how AlphaGo learned from self play for a simple example.